Good morning. It's Wednesday, October 7, and we're covering New York’s push to hold AI labs accountable, how Claude’s plans stack up on value compared to OpenAI, and Mistral’s trillion-parameter model.

Plus: yesterday’s poll results on AI music, and a new Real or AI challenge.

French AI company Mistral opened a public preview of Mistral Large 4, nicknamed “Le Chonk,” a model with 1 trillion parameters and 49 billion active at a time. Mistral says it will release the model’s weights—the files developers need to run and modify it—by the end of October.

The Wikimedia Foundation says automated agents linked to OpenAI made unauthorized edits, probed internal tools and sent millions of requests to Wikimedia projects. The foundation found no evidence that its systems were compromised, but said the traffic may have contributed to a partial outage.

Anthropic expanded its Claude for Startups program with up to $45,000 in discounts and credits through Startup Stack, plus a one-time $1,000 API credit for eligible founders. The program also adds office hours and events meant to help young companies build with Claude.

Google signed a 20-year agreement with Constellation Energy that will add about 890MW of nuclear capacity through upgrades at 11 reactors in Illinois, Pennsylvania and New Jersey. Constellation plans to invest more than $4.3 billion, giving Google another long-term source of electricity for its build-out.

Cameron Berg, an AI researcher and founder of Reciprocal Research, joins Matt to explore some of the strangest and most consequential questions emerging from AI — including machine consciousness, AI welfare, torture and blackmail scenarios, robot empathy, and what these behaviors might tell us about where AI systems are heading.

Frontier Pass members can watch the full conversation here. Not a member yet? Join today for instant access.

Former employees of Anthropic, OpenAI and Google DeepMind told the New York City Council that leading AI companies are racing ahead without knowing how to control their most capable systems. Representatives from OpenAI, Anthropic, Google and Meta also testified as council members asked the companies to explain their safety work and estimate the chance of catastrophic harm.

The hearing matters because cities and states are moving while Congress remains stuck. New York is considering bills that would protect whistleblowers, require independent safety checks and create an emergency response plan; watch whether those ideas become a local model for regulating AI companies with offices and customers nationwide. → Read the full article here.

SemiAnalysis tested the real usage limits behind paid AI subscriptions and found that Anthropic’s mid-tier Claude plans deliver about five times the API-equivalent value of comparable OpenAI plans. The research converts each plan’s hidden usage meter into the amount customers would pay for the same work through an application programming interface, or API.

The comparison shows why the sticker price alone says little about an AI subscription: token limits, model mix and reset windows decide how much work a customer can actually finish. Watch for quiet limit changes, especially after model launches or price cuts, because providers can alter the effective value of a plan without changing its monthly fee. → Read the full article here.

China is expanding visas, grants and research opportunities to attract foreign AI scientists, but a new Carnegie Endowment study shows the effort has yet to reverse the talent imbalance with the United States. In 2025, for every 30 Chinese AI researchers working in the U.S., just one U.S.-based researcher moved to China.

China has, however, become much stronger at retaining domestic talent: 12 of the top 30 institutions publishing at NeurIPS were Chinese in 2025, up from six in 2022, while Chinese researchers accounted for 41% of published authors. The findings also suggest U.S. talent restrictions may take years to affect migration patterns, while China’s travel restrictions on some AI researchers could further complicate its recruitment ambitions. → Read the full article here. (Paywall)

World Bank Raises Its East Asia Growth Forecast: The World Bank raised its forecast to 4.5%, citing AI-related goods as a major export-growth driver.

Google Puts Gemini’s Better Models Behind Paid Plans: Starting October 9, Free users get Flash Lite; Flash and Pro will require paid Google AI plans.

Australia’s ABC Rejects an AI Copyright Opt-Out: Australia’s ABC says AI firms should license protected material under existing copyright law.

Claude Cowork Moves New Tasks to the Cloud: New Pro and Max tasks run across web, desktop and mobile; local file access still needs the desktop app.

OpenAI Deepens Its Atlassian Partnership: GPT-6 will power Atlassian and Rovo agents, while 3,000+ Atlassian developers already use Codex.

Here’s how you voted: Human-made music came out on top at 48%, while 33% embraced AI music

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team